An Endpoint-Cloud Collaborative Intent-Driven AI Host System

Abstract

This invention discloses an edge-cloud collaborative intent-driven AI host system, relating to the fields of artificial intelligence and cloud computing technologies. It includes: an intent semantic tensor mapping module, which generates a three-dimensional semantic tensor through semantic parsing and structured transformation of multimodal inputs, achieving semantic disambiguation and structured expression of cross-modal information; a dynamic task offloading module, which dynamically optimizes the offloading path of tasks on the edge and cloud by constructing a Q-Learning reinforcement learning model, achieving intelligent resource allocation where low-computation tasks are executed locally and high-computation tasks are processed encrypted in the cloud; an edge-cloud collaborative architecture module, which achieves balanced computational load, low-latency response, and privacy protection through cloud-edge collaborative inference combined with differential privacy synchronization and real-time hot update mechanisms; and a privacy computing engine module, which adopts a dual-engine security architecture and achieves encrypted computing through the collaboration of homomorphic encryption and a trusted execution environment. This invention can achieve high-precision semantic parsing, low-latency response, and privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and cloud computing technology, specifically to an end-to-end cloud collaborative intent-driven AI host system. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI systems are widely used in various fields such as intelligent interaction, medical diagnosis, and smart homes. However, existing technologies have significant shortcomings in terms of intent parsing, real-time response, and privacy protection, making it difficult to meet the application needs in complex scenarios.

[0003] At the intent understanding level, traditional Natural Language Processing (NLP) models, such as Recurrent Neural Networks (RNNs) and their derivatives LSTM and GRU, primarily rely on single-modal text information for instruction parsing. Due to a lack of cross-modal associative capabilities, they cannot effectively integrate semantic features from different modalities when faced with complex instructions containing multimodal information such as images, speech, and text. For example, in intelligent customer service scenarios, when users simultaneously send voice commands and supplementary textual explanations, traditional RNN models struggle to integrate multimodal information for accurate understanding, resulting in a complex instruction parsing error rate exceeding 30%. This makes it difficult for AI systems to accurately execute user intent when handling complex tasks in real-world applications, severely impacting the user experience and task completion efficiency.

[0004] In terms of real-time response, existing AI systems mostly rely on centralized cloud computing. Taking the ResNet-50 model, commonly used in image recognition, as an example, due to its large number of parameters and high computational complexity, terminal devices need to upload data to cloud servers for processing before receiving the results returned from the cloud. This process results in an end-to-end latency exceeding 500ms. This latency cannot meet the real-time requirements of systems in scenarios with extremely high responsiveness, such as autonomous driving and real-time game interaction. For example, in autonomous driving scenarios, if the data collected by vehicle sensors undergoes a long delay in cloud processing, the vehicle will be unable to react to sudden road conditions in a timely manner, seriously threatening driving safety.

[0005] Privacy protection has become a major pain point for existing AI systems. When processing sensitive data such as biometrics (e.g., fingerprints, facial data) and medical images, traditional AI systems typically require users to upload the data in plaintext to the cloud for analysis and processing. This approach exposes the data to the risk of interception during transmission, and cloud storage also faces the risk of data leakage. For example, in the medical field, the leakage of sensitive information such as patients' CT images and genetic data would pose a serious threat to patients' privacy and security.

[0006] From a comparative perspective, while pure cloud-based solutions offer powerful computing resources and storage capabilities, they cannot guarantee real-time performance on the edge. For example, Google's federated learning framework, while achieving some level of data privacy protection, only supports small-scale models (1.4MB), failing to effectively support large-scale deep models required for complex tasks and thus unable to meet high-precision computing demands. Local compression solutions, such as the lightweight model TinyBERT, reduce computational and storage requirements through model compression, lowering reliance on cloud computing, but at the cost of semantic parsing accuracy, resulting in an F1-score drop to 82%, making it difficult to guarantee accuracy in practical applications. Furthermore, traditional encryption technologies, such as the Advanced Encryption Standard (AES), can only encrypt data for storage and transmission, not for encrypted computation. This makes it difficult to achieve compatibility between privacy protection and functionality in AI systems; encrypted data cannot be directly used for model computation and must be decrypted before processing, exposing data to privacy risks during the decryption phase.

[0007] In summary, the technical bottlenecks of existing AI systems in intent parsing, real-time response, and privacy protection limit their in-depth application in more critical fields. Therefore, there is an urgent need for an innovative technical solution to achieve high-precision semantic parsing, low-latency response, and end-to-end privacy and security, meeting the application needs of AI systems in complex scenarios. Summary of the Invention

[0008] This invention addresses the shortcomings of existing AI systems in intent parsing, real-time response, and privacy protection by providing an edge-cloud collaborative intent-driven AI host system to achieve high-precision semantic parsing, low-latency response, and end-to-end privacy and security.

[0009] The present invention provides an edge-cloud collaborative intent-driven AI host system, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0010] An edge-cloud collaborative intent-driven AI host system, the structure of which includes:

[0011] The Intent Semantic Tensor Mapping module generates a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints through semantic parsing and structured transformation of multimodal inputs. The output semantic tensor directly provides task attribute definitions for the dynamic task unloading module, realizing semantic disambiguation and structured expression of cross-modal information, and laying a semantic foundation for subsequent task allocation.

[0012] The dynamic task offloading module constructs a Q-Learning reinforcement learning model constrained by latency, privacy level, and energy consumption. Based on the semantic tensor output by the intent semantic tensor mapping module, it dynamically optimizes the task offloading path. Tasks that meet preset condition a are assigned to local execution, while tasks that meet preset condition b are encrypted and handed over to the edge-cloud collaborative architecture module for processing. At the same time, the privacy computing engine module is called to ensure data security. The offloading strategy is iterated based on the feedback results of the privacy computing engine to achieve intelligent resource allocation and ensure privacy security and energy efficiency.

[0013] The edge-cloud collaboration architecture module, based on the task unloading decision results of the dynamic task unloading module, achieves computational load balancing and low-latency response through collaborative reasoning between the cloud-based large model and the edge-side lightweight model, combined with differential privacy synchronization and real-time hot update mechanisms. When processing tasks that require privacy protection, the encryption capabilities of the privacy computing engine module are automatically triggered to ensure data security during the edge-cloud collaboration process.

[0014] The privacy computing engine module adopts a dual-engine security architecture. When the edge-cloud collaborative architecture module calls a sensitive data processing task, it achieves encrypted computing through the collaboration of homomorphic encryption and a trusted execution environment, and ensures the traceability of the operation through blockchain evidence storage. After the computing is completed, it feeds back the privacy protection status to the dynamic task unloading module, providing a basis for optimizing energy consumption and security level for subsequent task allocation strategies, forming an end-to-end closed loop for sensitive data security.

[0015] Optionally, the intent semantic tensor mapping module involved specifically includes:

[0016] The multi-source data fusion processing submodule is used to receive multimodal inputs of voice, text and sensor data, and convert unstructured data into a standardized sequence form. On the other hand, it processes long sequence dependencies based on the Transformer-XL architecture, captures long-distance associations of cross-modal information through an extended attention mechanism, and the output standardized sequence directly provides input with task attribute pre-annotation for the knowledge graph disambiguation submodule, laying the data foundation for the definition of task attributes in the subsequent semantic tensor.

[0017] The knowledge graph disambiguation submodule is used to perform semantic disambiguation by combining the domain knowledge graph. It uses a graph attention network to calculate the association weight between entities and graph nodes and dynamically selects high-confidence semantic paths. Domain logic constraints ensure the accuracy of entity and action semantics in the parsing results. The disambiguated semantic elements provide the tensor generation submodule with structured input with task attribute priority, avoiding policy deviation of the dynamic task unloading module due to semantic ambiguity.

[0018] The tensor generation submodule maps the parsed semantic elements into three-dimensional tensors containing entities, actions, and spatiotemporal constraints. The output tensors retain input information through cross-modal semantic associations and are then integrated into a structured JSON-LD format. The domain ontology is associated through context links to ensure that the generated semantic tensors can be directly parsed by the dynamic task unloading module as the basis for task allocation strategy decisions.

[0019] Preferably, the tensor generation submodule maps the parsed semantic elements into a three-dimensional tensor dimension containing entities, actions, and spatiotemporal constraints. This process specifically includes:

[0020] For the entity dimension, the subject and object are extracted from the input data and converted into low-dimensional semantic vectors through BERT word embeddings;

[0021] For the action dimension, the semantic behavioral instructions are parsed and combined with the domain action library to generate action codes;

[0022] For the spatiotemporal constraint dimension, time, space and condition constraints are extracted and converted into standardized timestamps, geographic coordinate vectors and Boolean condition matrices.

[0023] Optionally, the dynamic task unloading module involved specifically includes:

[0024] The multi-objective decision-making model submodule takes the semantic tensor output by the intent semantic tensor mapping module as input, parses the task attributes in it, and constructs the constraints of the Q-Learning reinforcement learning model by combining the real-time network status and device energy consumption status. On the other hand, it dynamically selects the task unloading path through the objective function of the Q-Learning reinforcement learning model and synchronizes the decision results to the task unloading execution submodule.

[0025] The task offloading and execution submodule performs preprocessing and executes tasks that meet preset condition a on the edge based on the results of the multi-objective decision model, ensuring local computing security through the trusted execution environment of the privacy computing engine module. On the other hand, it offloads tasks that meet preset condition b to the cloud, encrypts the data through homomorphic encryption of the privacy computing engine module, transmits it to the cloud, and calls the 3DResNet-152 model for inference.

[0026] The result feedback processing submodule receives the encrypted inference result returned from the cloud, calls the decryption capability of the privacy computing engine module to decrypt it on the edge, and displays it through a secure rendering channel; it feeds back the actual execution result to the Q-Learning reinforcement learning model, and dynamically adjusts the weight coefficients of the objective function based on the security audit logs provided by the privacy computing engine module; based on the updated weight coefficients, the Q-Learning reinforcement learning model re-determines the task unloading path, and synchronizes the optimized resource allocation strategy to the edge-cloud collaborative architecture module to guide it in adjusting the collaboration method between the large cloud model and the lightweight edge model.

[0027] Preferably, the objective function expression of the Q-Learning reinforcement learning model involved is:

[0028] Minimize: α*T+β*(1 / P)+γ*E,

[0029] In the formula, α, β, and γ are the weight coefficients of the initial configuration or the subsequent calculation and update.

[0030] Optionally, the specific edge-cloud collaborative architecture modules involved include:

[0031] The cloud-based submodule, relying on a GPU cluster to deploy a large GPT-4 model, serves as a deep processing center for multimodal data. It handles tasks assigned by the dynamic task offloading module, including: integrating cross-modal data of text, images, and videos, and outputting structured analysis results; aggregating differential privacy parameters uploaded from the client based on a federated learning framework, and periodically iterating and optimizing the global model; and processing tasks that require large-scale knowledge graph support, with encrypted results returned to the client.

[0032] The edge submodule runs the MobileBERT-Tiny lightweight model on the edge node, executes the tasks assigned by the dynamic task offloading module, and realizes the real-time execution of basic intents in offline scenarios, including: processing basic commands; performing real-time noise reduction and normalization on sensor data, generating floating-point feature vectors, and triggering encryption processing of the privacy computing engine according to the privacy level; parsing control commands and driving peripheral devices to execute, with the results directly fed back to the user;

[0033] The collaboration submodule, through differential privacy synchronization and real-time hot updates, dynamically balances model parameters and computational load between the cloud and the edge while protecting data privacy. This includes: receiving load balancing instructions from the dynamic task offloading module and adjusting the collaboration strategy between the large cloud model and the lightweight edge model; compressing edge model parameters using differential privacy technology and securely uploading them to the cloud for aggregation; and dynamically adjusting the accuracy and privacy protection strength of collaborative inference based on the security audit results of the privacy computing engine module through the real-time hot update mechanism.

[0034] Optionally, the privacy computing engine modules involved specifically include:

[0035] The homomorphic encryption submodule, based on the CKKS homomorphic encryption scheme, encrypts the floating-point feature vectors transmitted by the edge-cloud collaborative architecture module element by element to generate a ciphertext tensor structure; it supports matrix multiplication and nonlinear activation function approximation calculation in the encrypted state; the encryption process is constrained by the privacy level of the dynamic task unloading module, and the encrypted data is transmitted to the cloud through the TLS1.3 channel to ensure that the inference process is executed in the encrypted state throughout.

[0036] The Trusted Execution Submodule constructs an independent execution environment based on hardware-level security isolation technology, ensuring that sensitive data collected by the edge submodules directly enters the Trusted Execution Environment. It verifies the integrity of the environment through the Secure Boot mechanism combined with a dynamic root trust chain, and realizes full lifecycle management of keys, including key derivation based on HKDF, key storage based on TPM, and key rotation based on time factors. Based on the energy consumption constraints of the dynamic task offloading module, it selects the optimal decryption algorithm, and after calculation, feeds back the actual security level and energy consumption data to the dynamic task offloading module.

[0037] The blockchain evidence storage submodule, based on the Hyperledger Fabric consortium blockchain architecture, constructs an operation record evidence storage system. It receives decryption result metadata from the trusted execution submodule, signs it with ECDSA, writes it into the ledger, and organizes the metadata using a Merkle tree structure to support efficient integrity verification. It periodically provides security audit reports to the dynamic task unloading module as a basis for optimizing the task unloading strategy. It implements access control through smart contracts and automatically adjusts the evidence storage granularity according to the privacy level requirements of the dynamic task unloading module.

[0038] Further, optionally, the specific operations performed by the privacy computing engine module include:

[0039] In the data encryption and transmission stage, after the terminal submodule extracts the floating-point feature vector, the homomorphic encryption submodule encrypts the feature vector based on the CKKS scheme, and then transmits the encrypted data to the cloud through the TLS1.3 encryption channel; during this process, the dynamic task unloading module will constrain the encryption parameters according to the privacy level.

[0040] During the cloud-based encrypted computing phase, the cloud submodule performs deep learning model inference on the encrypted data without decryption. After the inference is completed, the result is returned to the edge in encrypted form. During this process, the trusted execution submodule monitors the energy consumption and security status and feeds back the relevant information to the dynamic task unloading module in real time.

[0041] In the edge-side decryption and evidence storage stage, the trusted execution submodule decrypts the encrypted inference results returned from the cloud in an independent execution environment. The metadata of the decrypted analysis results is then signed by the blockchain evidence storage submodule using the ECDSA elliptic curve signature algorithm and written into the Hyperledger Fabric ledger to form a traceable chain of inference records. During this process, the blockchain evidence storage submodule regularly provides security audit reports to the dynamic task unloading module, providing a basis for the dynamic task unloading module to optimize its task unloading strategy.

[0042] The edge-cloud collaborative intent-driven AI host system of the present invention has the following advantages compared with the prior art:

[0043] This invention can achieve high-precision semantic parsing, low-latency response, and end-to-end privacy and security; it is applicable to high-security scenarios such as medical diagnosis and financial transactions, and solves problems such as large intention understanding deviation, high edge latency, and privacy leakage risk in existing technologies, thereby improving system response efficiency and compliance.

[0044] This invention improves the parsing accuracy of cross-modal instructions and eliminates semantic ambiguity through an intent semantic tensor mapping module, reduces energy consumption through dynamic task offloading and modular resource allocation, reduces edge latency and balances computing load through an edge-cloud collaborative architecture module, and ensures end-to-end data security through a privacy computing engine module. Attached Figure Description

[0045] Appendix Figure 1 This is a block diagram of the module connection of the system described in the embodiment of the present invention. Detailed Implementation

[0046] To make the technical solution, the technical problem solved, and the technical effect of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments.

[0047] Example:

[0048] Reference Appendix Figure 1 This embodiment proposes an edge-cloud collaborative intent-driven AI host system, the structure of which includes:

[0049] The Intent Semantic Tensor Mapping module generates a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints through semantic parsing and structured transformation of multimodal inputs. The output semantic tensor directly provides task attribute definitions for the dynamic task unloading module, realizing semantic disambiguation and structured expression of cross-modal information, and laying a semantic foundation for subsequent task allocation.

[0050] The dynamic task offloading module constructs a Q-Learning reinforcement learning model constrained by latency, privacy level, and energy consumption. Based on the semantic tensor output by the intent semantic tensor mapping module, it dynamically optimizes the task offloading path. Tasks that meet preset condition a are assigned to local execution, while tasks that meet preset condition b are encrypted and handed over to the edge-cloud collaborative architecture module for processing. At the same time, the privacy computing engine module is called to ensure data security. The offloading strategy is iterated based on the feedback results of the privacy computing engine to achieve intelligent resource allocation and ensure privacy and energy efficiency.

[0051] The edge-cloud collaboration architecture module, based on the task unloading decision results of the dynamic task unloading module, achieves computational load balancing and low-latency response through collaborative reasoning between the cloud-based large model and the edge-side lightweight model, combined with differential privacy synchronization and real-time hot update mechanisms. When processing tasks that require privacy protection, the encryption capabilities of the privacy computing engine module are automatically triggered to ensure data security during the edge-cloud collaboration process.

[0052] The privacy computing engine module adopts a dual-engine security architecture. When the edge-cloud collaborative architecture module calls a sensitive data processing task, it achieves encrypted computing through the collaboration of homomorphic encryption and a trusted execution environment, and ensures the traceability of the operation through blockchain evidence storage. After the computing is completed, it feeds back the privacy protection status to the dynamic task unloading module, providing a basis for optimizing energy consumption and security level for subsequent task allocation strategies, forming an end-to-end closed loop for sensitive data security.

[0053] It should be added that a computational complexity threshold is preset. When the actual computational complexity threshold of a task is less than the preset threshold (i.e., the task meets preset condition a), the task is called a low-computation task and is subsequently assigned to local execution; when the actual computational complexity threshold of a task is greater than the preset threshold (i.e., the task meets preset condition b), the task is called a high-computation task and is subsequently offloaded to the cloud.

[0054] In this embodiment, the intent semantic tensor mapping module specifically includes:

[0055] The multi-source data fusion processing submodule is used to receive multimodal inputs of voice, text and sensor data, and convert unstructured data into a standardized sequence form. On the other hand, it processes long sequence dependencies based on the Transformer-XL architecture, captures long-distance associations of cross-modal information through an extended attention mechanism, and the output standardized sequence directly provides input with task attribute pre-annotation for the knowledge graph disambiguation submodule, laying the data foundation for the definition of task attributes in the subsequent semantic tensor.

[0056] The knowledge graph disambiguation submodule is used to perform semantic disambiguation by combining the domain knowledge graph. It uses a graph attention network to calculate the association weight between entities and graph nodes and dynamically selects high-confidence semantic paths. Domain logic constraints ensure the accuracy of entity and action semantics in the parsing results. The disambiguated semantic elements provide the tensor generation submodule with structured input with task attribute priority, avoiding policy deviation of the dynamic task unloading module due to semantic ambiguity.

[0057] The tensor generation submodule maps the parsed semantic elements into three-dimensional tensors containing entities, actions, and spatiotemporal constraints. The output tensors retain input information through cross-modal semantic association and are then integrated into a structured JSON-LD format. The domain ontology is associated through context links to ensure that the generated semantic tensors can be directly parsed by the dynamic task unloading module as the basis for task allocation strategy decisions.

[0058] The tensor generation submodule involved maps the parsed semantic elements into a three-dimensional tensor dimension containing entities, actions, and spatiotemporal constraints. This process specifically includes:

[0059] For the entity dimension, the subject and object are extracted from the input data and converted into low-dimensional semantic vectors through BERT word embeddings;

[0060] For the action dimension, the semantic behavioral instructions are parsed and combined with the domain action library to generate action codes;

[0061] For the spatiotemporal constraint dimension, time, space and condition constraints are extracted and converted into standardized timestamps, geographic coordinate vectors and Boolean condition matrices.

[0062] In this embodiment, the dynamic task unloading module specifically includes:

[0063] The multi-objective decision-making model submodule takes the semantic tensor output by the intent semantic tensor mapping module as input, parses the task attributes in it, and constructs the constraints of the Q-Learning reinforcement learning model by combining the real-time network status and device energy consumption status. On the other hand, it dynamically selects the task unloading path through the objective function of the Q-Learning reinforcement learning model and synchronizes the decision results to the task unloading execution submodule.

[0064] The task offloading and execution submodule performs preprocessing and executes tasks that meet preset condition a (i.e., low computational tasks) on the edge based on the results of the multi-objective decision model, ensuring local computation security through the trusted execution environment of the privacy computing engine module. On the other hand, it offloads tasks that meet preset condition b (i.e., high computational tasks) to the cloud, encrypts the data through homomorphic encryption of the privacy computing engine module, transmits it to the cloud, and calls the 3DResNet-152 model for inference.

[0065] The result feedback processing submodule receives the encrypted inference result returned from the cloud, calls the decryption capability of the privacy computing engine module to decrypt it on the edge, and displays it through a secure rendering channel; it feeds back the actual execution result to the Q-Learning reinforcement learning model, and dynamically adjusts the weight coefficients of the objective function based on the security audit logs provided by the privacy computing engine module; based on the updated weight coefficients, the Q-Learning reinforcement learning model re-determines the task unloading path, and synchronizes the optimized resource allocation strategy to the edge-cloud collaborative architecture module to guide it in adjusting the collaboration method between the large cloud model and the lightweight edge model.

[0066] The objective function expression for the Q-Learning reinforcement learning model involved is:

[0067] Minimize: α*T+β*(1 / P)+γ*E,

[0068] In the formula, α, β, and γ are the weight coefficients of the initial configuration or the subsequent calculation and update.

[0069] In this embodiment, the edge-cloud collaborative architecture module specifically includes:

[0070] The cloud-based submodule, relying on a GPU cluster to deploy a large GPT-4 model, serves as a deep processing center for multimodal data. It handles tasks assigned by the dynamic task offloading module, including: integrating cross-modal data of text, images, and videos, and outputting structured analysis results; aggregating differential privacy parameters uploaded from the client based on a federated learning framework, and periodically iterating and optimizing the global model; and processing tasks that require large-scale knowledge graph support, with encrypted results returned to the client.

[0071] The edge submodule runs the MobileBERT-Tiny lightweight model on the edge node, executes the tasks assigned by the dynamic task offloading module, and realizes the real-time execution of basic intents in offline scenarios, including: processing basic commands; performing real-time noise reduction and normalization on sensor data, generating floating-point feature vectors, and triggering encryption processing of the privacy computing engine according to the privacy level; parsing control commands and driving peripheral devices to execute, with the results directly fed back to the user;

[0072] The collaboration submodule, through differential privacy synchronization and real-time hot updates, dynamically balances model parameters and computational load between the cloud and the edge while protecting data privacy. This includes: receiving load balancing instructions from the dynamic task offloading module and adjusting the collaboration strategy between the large cloud model and the lightweight edge model; compressing edge model parameters using differential privacy technology and securely uploading them to the cloud for aggregation; and dynamically adjusting the accuracy and privacy protection strength of collaborative inference based on the security audit results of the privacy computing engine module through the real-time hot update mechanism.

[0073] In this embodiment, the privacy computing engine module specifically includes:

[0074] The homomorphic encryption submodule, based on the CKKS homomorphic encryption scheme, encrypts the floating-point feature vectors transmitted by the edge-cloud collaborative architecture module element by element to generate a ciphertext tensor structure; it supports matrix multiplication and nonlinear activation function approximation calculation in the encrypted state; the encryption process is constrained by the privacy level of the dynamic task unloading module, and the encrypted data is transmitted to the cloud through the TLS1.3 channel to ensure that the inference process is executed in the encrypted state throughout.

[0075] The Trusted Execution Submodule constructs an independent execution environment based on hardware-level security isolation technology, ensuring that sensitive data collected by the edge submodules directly enters the Trusted Execution Environment. It verifies the integrity of the environment through the Secure Boot mechanism combined with a dynamic root trust chain, and realizes full lifecycle management of keys, including key derivation based on HKDF, key storage based on TPM, and key rotation based on time factors. Based on the energy consumption constraints of the dynamic task offloading module, it selects the optimal decryption algorithm, and after calculation, feeds back the actual security level and energy consumption data to the dynamic task offloading module.

[0076] The blockchain evidence storage submodule, based on the Hyperledger Fabric consortium blockchain architecture, constructs an operation record evidence storage system. It receives decryption result metadata from the trusted execution submodule, signs it with ECDSA, writes it into the ledger, and organizes the metadata using a Merkle tree structure to support efficient integrity verification. It periodically provides security audit reports to the dynamic task unloading module as a basis for optimizing the task unloading strategy. It implements access control through smart contracts and automatically adjusts the evidence storage granularity according to the privacy level requirements of the dynamic task unloading module.

[0077] The specific operations performed by the privacy computing engine module include:

[0078] In the data encryption and transmission stage, after the terminal submodule extracts the floating-point feature vector, the homomorphic encryption submodule encrypts the feature vector based on the CKKS scheme, and then transmits the encrypted data to the cloud through the TLS1.3 encryption channel; during this process, the dynamic task unloading module will constrain the encryption parameters according to the privacy level.

[0079] During the cloud-based encrypted computing phase, the cloud submodule performs deep learning model inference on the encrypted data without decryption. After the inference is completed, the result is returned to the edge in encrypted form. During this process, the trusted execution submodule monitors the energy consumption and security status and feeds back the relevant information to the dynamic task unloading module in real time.

[0080] In the edge-side decryption and evidence storage stage, the trusted execution submodule decrypts the encrypted inference results returned from the cloud in an independent execution environment. The metadata of the decrypted analysis results is then signed by the blockchain evidence storage submodule using the ECDSA elliptic curve signature algorithm and written into the Hyperledger Fabric ledger to form a traceable chain of inference records. During this process, the blockchain evidence storage submodule regularly provides security audit reports to the dynamic task unloading module, providing a basis for the dynamic task unloading module to optimize its task unloading strategy.

[0081] A detailed description of an embodiment based on the medical image diagnosis workflow, including the technical implementation and interaction logic of the modules described in this embodiment:

[0082] 1. The doctor gives verbal instructions, which are then converted into text by the speech recognition module and transmitted to the intent semantic tensor mapping module;

[0083] 2. In the intent semantic tensor mapping module, the multi-source data fusion processing submodule uses Transformer-XL to extract keywords and convert them into standardized sequences with task attributes; the knowledge graph disambiguation submodule associates with the SNOMED-CT atlas and generates high-confidence semantic paths after disambiguation; the tensor generation submodule constructs a three-dimensional semantic tensor containing entities, actions and spatiotemporal constraints, converts it into JSON-LD and associates it with the medical ontology for decision-making by the dynamic task unloading module;

[0084] 3. The dynamic task unloading module parses semantic tensors and combines them with client-side resources.

[0085] i) Decision-making low computational tasks (such as CT image window width adjustment, ROI segmentation) are executed on the edge: instructions are sent to the edge-cloud collaborative architecture module, the edge sub-module of the edge-cloud collaborative architecture module adjusts the CT window width in an independent execution environment, segments nodules with lightweight MobileBERT-Tiny, generates feature vectors and sends them to the privacy computing engine module for encryption; the edge sub-module reports the resource consumption to the dynamic task offloading module.

[0086] ii) Offloading computationally intensive decision-making tasks (such as malignancy classification and gene sequence analysis) to the cloud: The offloading path is optimized through the Q-Learning reinforcement learning model. The feature vector is encrypted by the homomorphic encryption submodule (CKKS scheme) of the privacy computing engine and transmitted to the cloud submodule of the edge-cloud collaborative architecture module through the TLS1.3 channel. The cloud submodule relies on the GPT-4 large model deployed on the GPU cluster to perform deep learning inference in the encrypted state and generate malignancy probability prediction results. After the cloud inference is completed, the encrypted result is returned to the edge side, which is received by the edge submodule of the edge-cloud collaborative architecture module and simultaneously feeds back the cloud computing resource usage (such as GPU time consumption and network latency) and encryption transmission status to the dynamic task offloading module.

[0087] 4. The homomorphic encryption submodule of the privacy computing engine module uses CKKS to encrypt feature vectors and transmits them to the cloud submodule via TLS1.3; the cloud submodule uses the GPT-4 large model to perform encrypted inference, generate probabilities and add differential privacy noise; the blockchain evidence storage submodule of the privacy computing engine module signs the inference metadata and writes it into the Hyperledger Fabric ledger; the cloud submodule reports the cloud resource consumption to the dynamic task unloading module.

[0088] 5. The trusted execution submodule of the privacy computing engine module decrypts the results in an independent execution environment, verifies their integrity, and generates diagnostic suggestions; the edge-side submodule of the edge-cloud collaborative architecture module displays the results through a secure rendering channel and transmits diagnostic metadata to the blockchain for evidence storage; the blockchain evidence storage submodule uploads the results to the blockchain, generates a hash value, and synchronizes it to the dynamic task unloading module; the trusted execution submodule feeds back the decryption energy consumption and security level to the dynamic task unloading module for optimizing subsequent strategies.

[0089] In summary, the edge-cloud collaborative intent-driven AI host system of the present invention can achieve high-precision semantic parsing, low-latency response, and end-to-end privacy and security, solving problems such as large intent understanding deviation, high edge latency, and privacy leakage risk in the prior art.

[0090] The above specific examples illustrate the principles and implementation methods of the present invention in detail. These embodiments are merely for the purpose of helping to understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made to the present invention by those skilled in the art without departing from the principles of the present invention should fall within the patent protection scope of the present invention.

Claims

1. An end-to-end cloud collaborative intent-driven AI host system, characterized in that, Its structure includes: The Intent Semantic Tensor Mapping module generates a three-dimensional semantic tensor containing entities, actions, and spatiotemporal constraints through semantic parsing and structured transformation of multimodal inputs. The output semantic tensor directly provides task attribute definitions for the dynamic task unloading module, realizing semantic disambiguation and structured expression of cross-modal information, and laying a semantic foundation for subsequent task allocation. The dynamic task offloading module constructs a Q-Learning reinforcement learning model constrained by latency, privacy level, and energy consumption. Based on the semantic tensor output by the intent semantic tensor mapping module, it dynamically optimizes the task offloading path. Tasks that meet preset condition a are assigned to local execution, while tasks that meet preset condition b are encrypted and handed over to the edge-cloud collaborative architecture module for processing. At the same time, the privacy computing engine module is called to ensure data security. The offloading strategy is iterated based on the feedback results of the privacy computing engine to achieve intelligent resource allocation and ensure privacy security and energy efficiency. The edge-cloud collaboration architecture module, based on the task unloading decision results of the dynamic task unloading module, achieves computational load balancing and low-latency response through collaborative reasoning between the cloud-based large model and the edge-side lightweight model, combined with differential privacy synchronization and real-time hot update mechanisms. When processing tasks that require privacy protection, the encryption capabilities of the privacy computing engine module are automatically triggered to ensure data security during the edge-cloud collaboration process. The privacy computing engine module adopts a dual-engine security architecture. When the edge-cloud collaborative architecture module calls a sensitive data processing task, it achieves encrypted computing through the collaboration of homomorphic encryption and a trusted execution environment, and ensures the traceability of the operation through blockchain evidence storage. After the calculation is completed, it feeds back the privacy protection status to the dynamic task unloading module, providing a basis for optimizing energy consumption and security level for subsequent task allocation strategies, forming an end-to-end closed loop for sensitive data security. Specifically, the dynamic task unloading module includes: The multi-objective decision-making model submodule takes the semantic tensor output by the intent semantic tensor mapping module as input, parses the task attributes in it, and constructs the constraints of the Q-Learning reinforcement learning model by combining the real-time network status and device energy consumption status. On the other hand, it dynamically selects the task unloading path through the objective function of the Q-Learning reinforcement learning model and synchronizes the decision results to the task unloading execution submodule. The task offloading and execution submodule performs preprocessing and executes tasks that meet preset condition a on the edge based on the results of the multi-objective decision model, ensuring local computing security through the trusted execution environment of the privacy computing engine module. On the other hand, it offloads tasks that meet preset condition b to the cloud, encrypts the data through homomorphic encryption of the privacy computing engine module, transmits it to the cloud, and calls the 3DResNet-152 model for inference. The result feedback processing submodule receives the encrypted inference result returned from the cloud, calls the decryption capability of the privacy computing engine module to decrypt it on the edge, and displays it through a secure rendering channel; it feeds back the actual execution result to the Q-Learning reinforcement learning model, and dynamically adjusts the weight coefficients of the objective function in combination with the security audit logs provided by the privacy computing engine module; based on the updated weight coefficients, the Q-Learning reinforcement learning model re-determines the task unloading path, and synchronizes the optimized resource allocation strategy to the edge-cloud collaborative architecture module to guide it in adjusting the collaboration method between the large cloud model and the lightweight edge model; The privacy computing engine module specifically includes: The homomorphic encryption submodule, based on the CKKS homomorphic encryption scheme, encrypts the floating-point feature vectors transmitted by the edge-cloud collaborative architecture module element by element to generate a ciphertext tensor structure; it supports matrix multiplication and nonlinear activation function approximation calculation in the encrypted state; the encryption process is constrained by the privacy level of the dynamic task unloading module, and the encrypted data is transmitted to the cloud through the TLS 1.3 channel to ensure that the inference process is executed in the encrypted state throughout; The Trusted Execution Submodule constructs an independent execution environment based on hardware-level security isolation technology, ensuring that sensitive data collected by the edge submodules directly enters the Trusted Execution Environment. It verifies the integrity of the environment through the Secure Boot mechanism combined with a dynamic root trust chain, and realizes full lifecycle management of keys, including key derivation based on HKDF, key storage based on TPM, and key rotation based on time factors. Based on the energy consumption constraints of the dynamic task offloading module, it selects the optimal decryption algorithm, and after calculation, feeds back the actual security level and energy consumption data to the dynamic task offloading module. The blockchain evidence storage submodule, based on the Hyperledger Fabric consortium blockchain architecture, constructs an operation record evidence storage system. It receives decryption result metadata from the trusted execution submodule, signs it with ECDSA, writes it into the ledger, and organizes the metadata using a Merkle tree structure to support efficient integrity verification. It periodically provides security audit reports to the dynamic task unloading module as a basis for optimizing the task unloading strategy. It implements access control through smart contracts and automatically adjusts the evidence storage granularity according to the privacy level requirements of the dynamic task unloading module.

2. The edge-cloud collaborative intent-driven AI host system according to claim 1, characterized in that, The intent semantic tensor mapping module specifically includes: The multi-source data fusion processing submodule is used to receive multimodal inputs of voice, text and sensor data, and convert unstructured data into a standardized sequence form. On the other hand, it processes long sequence dependencies based on the Transformer-XL architecture, captures long-distance associations of cross-modal information through an extended attention mechanism, and the output standardized sequence directly provides input with task attribute pre-annotation for the knowledge graph disambiguation submodule, laying the data foundation for the definition of task attributes in the subsequent semantic tensor. The knowledge graph disambiguation submodule is used to perform semantic disambiguation by combining the domain knowledge graph. It uses a graph attention network to calculate the association weight between entities and graph nodes and dynamically selects high-confidence semantic paths. Domain logic constraints ensure the accuracy of entity and action semantics in the parsing results. The disambiguated semantic elements provide the tensor generation submodule with structured input with task attribute priority, avoiding policy deviation of the dynamic task unloading module due to semantic ambiguity. The tensor generation submodule maps the parsed semantic elements into three-dimensional tensors containing entities, actions, and spatiotemporal constraints. The output tensors retain input information through cross-modal semantic associations and are then integrated into a structured JSON-LD format. The domain ontology is associated through context links to ensure that the generated semantic tensors can be directly parsed by the dynamic task unloading module as the basis for task allocation strategy decisions.

3. The edge-cloud collaborative intent-driven AI host system according to claim 2, characterized in that, The tensor generation submodule maps the parsed semantic elements into a three-dimensional tensor dimension that includes entities, actions, and spatiotemporal constraints. This process specifically includes: For the entity dimension, the subject and object are extracted from the input data and converted into low-dimensional semantic vectors through BERT word embeddings; For the action dimension, the semantic behavioral instructions are parsed and combined with the domain action library to generate action codes; For the spatiotemporal constraint dimension, time, space and condition constraints are extracted and converted into standardized timestamps, geographic coordinate vectors and Boolean condition matrices.

4. The edge-cloud collaborative intent-driven AI host system according to claim 1, characterized in that, The objective function expression of the Q-Learning reinforcement learning model is: Minimize: α*T+β*(1 / P)+γ*E, In the formula, α, β, and γ are the weight coefficients of the initial configuration or the subsequent calculation and update.

5. The edge-cloud collaborative intent-driven AI host system according to claim 1, characterized in that, The edge-cloud collaborative architecture module specifically includes: The cloud-based submodule, relying on a GPU cluster to deploy a large GPT-4 model, serves as a deep processing center for multimodal data. It handles tasks assigned by the dynamic task offloading module, including: integrating cross-modal data of text, images, and videos, and outputting structured analysis results; aggregating differential privacy parameters uploaded from the client based on a federated learning framework, and periodically iterating and optimizing the global model; and processing tasks that require large-scale knowledge graph support, with encrypted results returned to the client. The edge submodule runs the MobileBERT-Tiny lightweight model on the edge node, executes the tasks assigned by the dynamic task offloading module, and realizes the real-time execution of basic intents in offline scenarios, including: processing basic commands; performing real-time noise reduction and normalization on sensor data, generating floating-point feature vectors, and triggering encryption processing of the privacy computing engine according to the privacy level; parsing control commands and driving peripheral devices to execute, with the results directly fed back to the user; The collaboration submodule, through differential privacy synchronization and real-time hot updates, dynamically balances model parameters and computational load between the cloud and the edge while protecting data privacy. This includes: receiving load balancing instructions from the dynamic task offloading module and adjusting the collaboration strategy between the large cloud model and the lightweight edge model; compressing edge model parameters using differential privacy technology and securely uploading them to the cloud for aggregation; and dynamically adjusting the accuracy and privacy protection strength of collaborative inference based on the security audit results of the privacy computing engine module through the real-time hot update mechanism.

6. The edge-cloud collaborative intent-driven AI host system according to claim 1, characterized in that, The specific operations performed by the privacy computing engine module include: In the data encryption and transmission stage, after the terminal submodule extracts the floating-point feature vector, the homomorphic encryption submodule encrypts the feature vector based on the CKKS scheme, and then transmits the encrypted data to the cloud through the TLS 1.3 encryption channel; during this process, the dynamic task unloading module will constrain the encryption parameters according to the privacy level. During the cloud-based encrypted computing phase, the cloud submodule performs deep learning model inference on the encrypted data without decryption. After the inference is completed, the result is returned to the edge in encrypted form. During this process, the trusted execution submodule monitors the energy consumption and security status and feeds back the relevant information to the dynamic task unloading module in real time. In the edge-side decryption and evidence storage stage, the trusted execution submodule decrypts the encrypted inference results returned from the cloud in an independent execution environment. The metadata of the decrypted analysis results is then signed by the blockchain evidence storage submodule using the ECDSA elliptic curve signature algorithm and written into the Hyperledger Fabric ledger to form a traceable chain of inference records. During this process, the blockchain evidence storage submodule regularly provides security audit reports to the dynamic task unloading module, providing a basis for the dynamic task unloading module to optimize its task unloading strategy.

Images